AgentSkillsCN

grounded-query

根据原始文件核实事实性陈述。当用户就会议内容、决策事项、项目细节，或任何需以数据文件为依据的主张进行提问时，可调用此技能。

SKILL.md
--- frontmatter
name: grounded-query
description: Verifies factual claims against source documents. Use when answering questions about meeting content, decisions, project details, or any assertion that should be grounded in data files.

Purpose

Ensure all factual claims in responses are supported by evidence from source documents. Prevents hallucination and unsupported status claims.

When to Use

Automatically apply this workflow when:

  • •Answering questions about what was discussed, decided, or planned
  • •Making claims about project status, timelines, or deliverables
  • •Summarizing meeting content or extracting information from transcripts
  • •Any query where accuracy matters and source files exist

Workflow

Step 1: Generate Initial Response

Answer the user's query based on your understanding of the data files.

Step 2: Extract Verifiable Claims

Identify statements in your response that are factual assertions, such as:

  • •"The team decided to use X"
  • •"Feature Y is complete"
  • •"Person Z mentioned..."
  • •Dates, names, decisions, status claims

Step 3: Search for Evidence

For each claim, search the data directory for supporting evidence:

  • •Use Grep to find relevant mentions
  • •Use Read to examine context around matches
  • •Note the file path and relevant line/section

Step 4: Classify Evidence

For each claim, assign a status:

  • •SUPPORTED - Direct evidence found in source files
  • •PARTIAL - Related mentions but not exact confirmation
  • •NOT FOUND - No evidence located (doesn't mean false, just unverified)

Step 5: Output Format

Always end your response with an evidence table:

code
### Evidence

| Claim | Status | Source |
|-------|--------|--------|
| [Brief claim] | SUPPORTED/PARTIAL/NOT FOUND | folder/filename |

If evidence is NOT FOUND for key claims, add a note:

Note: Some claims could not be verified against available data files. Confidence is lower for unverified items.

Example

Query: "What database did the team choose?"

Response: The team chose Supabase as the primary database, with pgvector for embeddings support.

Evidence

ClaimStatusSource
Supabase as primary databaseSUPPORTEDgranola/2025-12-22T02_50_00Z.md
pgvector for embeddingsSUPPORTEDapple-note-voice-memos/2025-12-21T04_48_00Z.md

Important

  • •Do NOT claim something is "Done" or "In Progress" based only on meeting discussions
  • •Distinguish between "discussed/triage" and "implemented"
  • •When in doubt, mark as PARTIAL or NOT FOUND rather than SUPPORTED

After Output

This skill returns results to the calling context. Do not stop execution. Continue with the next step in the workflow or TODO list.